JSON Repair — Fix Broken JSON, LLM Output & Trailing Commas

What is JSON Repair?

A JSON repair tool takes text that is 'almost JSON' and rewrites it into strictly valid JSON. It became a critical piece of the LLM stack because language models routinely emit JSON wrapped in markdown fences, sprinkled with trailing commas, or seasoned with Python-style literals like True and None. Standard JSON.parse rejects all of those. Rather than manually strip fences and re-quote keys every time, this tool applies a well-defined chain of transformations — strip fences, normalize smart quotes, extract the first balanced object/array, then lex-scan to remove comments, quote bare keys, drop trailing commas, and coerce special literals — and re-serializes the result through JSON.stringify so the output is guaranteed to be spec-compliant.

How to repair broken JSON online

1. 1Paste the broken JSON — LLM output, JSON5, JSONC, or a hand-edited config — into the left editor.
2. 2The right pane instantly shows the repaired JSON, formatted with 2-space indent.
3. 3Review the list of applied fixes in the status card. Every transformation is disclosed by name (trailing commas removed, single quotes converted, etc.).
4. 4If the input can't be salvaged, the tool shows a parser error with line and column pointing at the first unrecoverable syntax problem, plus the partial fixes it did manage to apply.
5. 5Copy the repaired output or switch to Tree view for a structural preview before shipping.

Use Cases

Code Examples

```json
{
  // The user's profile
  name: 'Alice',
  admin: True,
  tags: ['dev', 'design',],
  score: NaN,
}
```

{
  "name": "Alice",
  "admin": true,
  "tags": [
    "dev",
    "design"
  ],
  "score": null
}

import { repairJson } from "./repair.js";

const raw = await callLLM({ prompt });
const { ok, output, patches } = repairJson(raw);
if (!ok) throw new Error("LLM output unrecoverable");
const payload = JSON.parse(output);
console.log(`repaired with ${patches.length} fixes`);

Key Concepts

RFC 8259

The current JSON specification. Requires double-quoted strings, no trailing commas, no comments, and no NaN/Infinity. Everything the repair tool touches is a deviation from this spec.

JSON5

A superset of JSON that allows single quotes, unquoted keys, comments, trailing commas, and NaN/Infinity. The repair tool essentially converts JSON5 to strict JSON, losing only the special float values (coerced to null).

JSONC

'JSON with Comments' — a strict-JSON superset used by VS Code (tsconfig.json, .vscode/settings.json) that permits // and /* */ comments. Stripping comments is one of the tool's core steps.

Structured output

Modern LLM APIs expose 'JSON mode' or 'response_format' options that constrain output to valid JSON. When those aren't available or return malformed data, this tool acts as a fallback layer before parsing.

Tips & Best Practices

• ▸Never trust LLM output blindly — after repair, still validate against a JSON Schema for the fields you actually care about.
• ▸If you control the prompt, ask the model to output raw JSON without markdown fences. Repair is a safety net, not a workaround for weak prompting.
• ▸The repair tool coerces NaN and Infinity to null. If those values are semantically important, encode them as strings ("NaN") in your schema instead.
• ▸Bare-key detection only fires when the identifier is immediately followed by a colon in an object context. Standalone identifiers like `true` / `false` / `null` are preserved as JSON literals.
• ▸For very large LLM outputs, batch through the tool in server-side JavaScript rather than the browser to avoid main-thread jank.